Research map / Label-efficient and robust learning / OOD and anomaly detection
What If the Input is Expanded in OOD Detection?
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CoVer, which averages OOD confidence scores over the original input and many common-corruption variants, exploiting a "confidence mutation" effect to improve separation of ID and OOD data for both traditional and CLIP-based zero-shot OOD detection.
Problem
Existing OOD scoring functions use only a single input, which limits the representation dimension available for separating hard OOD samples from in-distribution data.
Approach
Expand the input dimension by applying common image corruptions and average the OOD scores over the original and corrupted inputs; overconfident OOD samples lose confidence under corruption much more than confident ID samples ("confidence mutation"), so the averaged score separates the distributions better.
Key results
- On ImageNet-1K with CLIP-B/16, CoVer reaches 92.45 average AUROC and 34.88 FPR95 versus 90.76 / 42.73 for MCM.
- With ResNet-50, ASH-S+CoVer improves average FPR95 from 22.74 to 21.83 and ASH-B+CoVer improves it from 39.04 to 26.53.
- CoVer improves MCM, LoCoOp, CLIPN and NegLabel across most OOD sets, e.g. MCM average AUROC 90.76 to 92.10.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Average confidence scores from original and corrupted inputs rather than using MCM on one input | CLIP-B/16 · ImageNet-1K ID, ImageNet OOD average | AUROC: 90.76 -> 92.45 | Helped |
| Add CoVer score averaging to ASH-S | ResNet-50 · ImageNet-1K ID, ImageNet OOD average | AUROC: 95.14 -> 95.40 | Helped |
| Add CoVer score averaging to ReAct | ResNet-50 · ImageNet-1K ID, ImageNet OOD average | AUROC: 92.95 -> 94.48 | Helped |
| Add CoVer score averaging to MCM | CLIP-B/16 · ImageNet-1K ID, ImageNet OOD average | AUROC: 90.76 -> 92.10 | Helped |
| Add CoVer score averaging to NegLabel | CLIP-B/16 · ImageNet-1K ID, ImageNet OOD average | AUROC: 94.10 -> 95.39 | Helped |
| Use original-plus-corrupted inputs instead of only a brightness-corrupted input (ablation) | CLIP-B/16 CoVer · ImageNet-1K ID, ImageNet OOD average | AUROC: 89.21 -> 91.5 | Helped |
| Use original-plus-corrupted inputs instead of only a contrast-corrupted input (ablation) | CLIP-B/16 CoVer · ImageNet-1K ID, ImageNet OOD average | AUROC: 74.2 -> 91.77 | Helped |
| Use original-plus-corrupted inputs instead of only a motion-blurred input (ablation) | CLIP-B/16 CoVer · ImageNet-1K ID, ImageNet OOD average | AUROC: 74.65 -> 90.35 | Helped |
| Use original-plus-corrupted inputs instead of only a spatter-corrupted input (ablation) | CLIP-B/16 CoVer · ImageNet-1K ID, ImageNet OOD average | AUROC: 76.86 -> 89.43 | Helped |
| Increase expanded scoring dimensions (ablation) | CoVer · OOD detection benchmarks | OOD detection performance | Mixed |
| Use severe corruptions such as Spatter or Elastic transform (ablation) | CoVer · OOD detection benchmarks | OOD detection performance | Hurt |
Surprising findings
- Single corrupted inputs usually degrade ID-OOD separability, but averaging over original and corrupted inputs improves it.
- Severe corruptions such as Spatter and Elastic transform can make CoVer worse than the single-input baseline.
Limitations
- Performance depends on choosing corruption types and severity, which is tuned on SVHN as a validation set with no standardized criterion.
- Severe corruptions (e.g., Spatter, Elastic transform) can hurt performance below the single-input baseline.
- Multiple forward passes per input increase inference cost.
Experimental setup
Datasets: ImageNet-1K SVHN iNaturalist SUN Places Textures ImageNet-10 ImageNet-20 ImageNet-100 WaterBirds
Benchmarks: ImageNet-1K OOD benchmark (iNaturalist, SUN, Places, Textures) ImageNet-10/20/100 hard OOD WaterBirds spurious OOD
Metrics: AUROC FPR95
Baselines: MSP ODIN Mahalanobis Energy ReAct DICE ASH MCM GradNorm ViM KNN ZOC
Builds on: MCM ASH ReAct DICE NegLabel
Closest papers by meaning
- The Best of Both Worlds: On the Dilemma of Out-of-distribution Detection (NeurIPS 2024 · similarity 0.90)
- Test-Time Linear Out-of-Distribution Detection (CVPR 2024 · similarity 0.90)
- Expecting The Unexpected: Towards Broad Out-Of-Distribution Detection (NeurIPS 2024 · similarity 0.89)
- Enhancing the Power of OOD Detection via Sample-Aware Model Selection (CVPR 2024 · similarity 0.89)
- Position: Supervised Classifiers Answer the Wrong Questions for OOD Detection (ICML 2025 · similarity 0.89)
- Out-of-Distribution Detection with Prototypical Outlier Proxy (AAAI 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.